Conversation

Notices

1. Embed this notice
   allison (aparrish@friend.camp)'s status on Thursday, 20-Aug-2026 07:14:33 JST allison

   the essential limitation with llm-generated code doesn't have anything to do with the capabilities of the model imo. the limitation is epistemological: how do I know if the code is correct? in order to know that for sure, you have to understand what the code is doing. in order to understand what the code is doing, you need to exert at least as much effort as you would have writing the code yourself in the first place. so the llm generation is always strictly a waste

   • Embed this notice
     Rene Wooller (renetron@aus.social)'s status on Thursday, 20-Aug-2026 08:05:15 JST Rene Wooller

     in reply to

     • Thomas 🔭🕹️

     @aparrish @thomasfuchs we used to have holy wars about things like tabs VS spaces. People would have long eloquent arguments as to why it should be one or the other. I think we’re well and truly past that (I mean spaces win obviously) or at least I don’t need to care about that kind of detail LLM writes it and it gets linted. Is this the right way to solve the problem is something you still have to think about but the implementation detail less so. Interestingly, strongly typed languages are becoming relatively more effective as the barrier to getting it to compile has diminished to almost nothing, while retaining the benefits

   • Embed this notice
     allison (aparrish@friend.camp)'s status on Thursday, 20-Aug-2026 08:05:16 JST allison

     • Thomas 🔭🕹️

     @thomasfuchs i understand what you're getting at here and i'm sympathetic to it, but code isn't just math, and ultimately i don't think its functioning is any less dependent on context than any other form of language. when i'm talking about correctness, i'm not (just) talking about formal correctness; i'm asking *is this the right way to solve the problem*, which, in the real world code functions in, is not a question that has a formal (or even statistical) answer